Graph neural networks: A paradigm shift in static signature classification via Delaunay triangulation

Abstract

Handwritten static signatures are a fundamental biometric characteristic extensively utilized across various sectors such as finance, law, and business. Graph neural networks (GNNs) have emerged as a promising avenue, offering inherent advantages in processing graph-based data. Inspired by the success of neural networks, the present investigation centers on harnessing the potential of GNNs for graph classification, particularly in the domain of static signature verification. The present study introduces a novel end-to-end methodology named Static Signature Verification using Delaunay Triangulation-based Graph Neural Network (DTGNN). This innovative approach involves converting static signature images into graph-structured data while integrating position embeddings to effectively preserve spatial information. By leveraging the capabilities of GNNs, especially in aggregating graph information, the approach aims to enhance the performance of static signature verification systems. It introduced novel Delaunay Triangulation-based Features for efficient model training, facilitating enhanced accuracy and reliability. The experimental assessments carried out on the GPDSsynthetic_100 and MCYT_75 datasets reveal notable improvements in performance, characterized by increased accuracy (GPDSsynthetic_100 - 99.91%; MCYT_75 - 100 %) and minimal loss (GPDSsynthetic_100 & MCYT_75 - 0.0012) utilizing our proposed approach. These results underscore the potential of DTGNN in addressing real-world verification challenges, highlighting its efficacy in writer-independent and writer-dependent static signature verification scenarios.